What's new in Mirakash

New

• Escalation queue — when an agent reaches its limits it opens an escalation; a live inbox lets human experts pick it up, resolve, and hand control back to the caller without ever dropping the line.

• Consult-and-resume — the agent side-calls a human, a phone number, or another AI agent, gets the answer, and returns to the conversation instead of transferring away. Agent-to-agent consults resolve live; person and number targets resolve asynchronously.

• Expert routing — assign escalations to providers by skill and availability; assignment, resolution, and hand-back are tracked end to end on the conversation record.

• Avatar gateway — a pluggable talking-head layer with a reference adapter; the agent's audio drives a face for web-chat and kiosk deployments, falling back to a reactive placeholder when no vendor is configured.

• Live conversation monitor — watch active voice and chat sessions in real time from the operator console, with per-stage state and running cost.

• Guardrail hand-off — when a guardrail trips mid-conversation, the agent can escalate automatically so risky turns reach a human immediately.

• Conversation summaries — every session closes with an auto-generated summary, outcome, and next action written back to the record.

Improved

• Escalations, live monitoring, and avatar controls are all permission- and feature-gated across the operator console and agent studio.

• Consult and escalation events flow onto the same unified conversation timeline as calls and chats, so no handoff is ever a black box.

New

• Omnichannel deployment — publish one agent to voice, WhatsApp, SMS, email, and a drop-in web chat widget from a single configuration.

• White-labeled web chat — an embeddable chat widget you brand as your own and drop onto any site with one snippet.

• Channel routing — inbound messages and calls route to the right agent and channel automatically.

• Unified inbox — voice and chat conversations land in one operator view with a shared transcript format.

Changed

• Conversations now carry an explicit channel, so transcripts, analytics, and costs can never mix across voice and chat.

New

• Containment & escalation analytics — see the share of conversations an agent resolves on its own versus hands off to a human.

• Sentiment tracking — per-conversation sentiment scored across the call and surfaced on the timeline and in reports.

• Latency dashboards — per-stage STT, LLM, TTS, and transport timing charted across every agent.

• Itemized cost per agent — STT, LLM, TTS, and transport costs broken out per agent and per minute, so unit economics are never a guess.

New

• LLM-judge scorecards — every conversation scored automatically against a rubric you define.

• Flagging rules & issues — write rules that surface risky or failed conversations; a failure opens a tracked issue.

• Review queue — a human review workflow for flagged conversations, with dispositions rolled back into scoring.

New

• Connectors — real, environment-gated integrations for Slack, Zendesk, Google Sheets, Notion, and outbound webhooks.

• Inbound webhooks & connection tests — external systems can start a workflow, and every connector ships with a one-click credential test.

• Copilot — a page-aware assistant sidebar that answers from your own workspace data and proposes actions, scoped to the caller's permissions.

• Environment variables — connector base URLs, keys, and endpoints are stored as {{VAR}} references and resolved per environment at runtime.

New

• One-click number provisioning — buy and attach a Twilio phone number to an agent across three account models: bring-your-own, a managed subaccount we run for you, or the platform master line.

• Encrypted carrier credentials — auth tokens and secrets are AES-256-GCM encrypted at rest and decrypted only at call time.

Changed

• Telephony credentials moved into the encrypted store; legacy plaintext configuration is migrated automatically and retired.

New

• Knowledge bases — upload documents and URLs; content is chunked and embedded automatically for retrieval.

• Retrieval-augmented answers — agents ground their replies in your knowledge at conversation time, with citations back to the source.

• Notion connector — sync Notion pages into a knowledge base and keep them current as they change.

• Per-agent scoping — attach specific knowledge bases to specific agents so each one only knows what it should.

New

• Visual workflow builder — an event-driven node graph with tools, conditions, transfer-to-agent, and transfer-to-human, validated live as you build.

• Tool nodes — call external APIs and built-in actions inline during a conversation and branch on the result.

• Runtime variable interpolation — {{VAR}} tokens in endpoints, headers, and arguments resolve from environment and conversation context at execution time.

• Run history — every workflow run logged step by step and inspectable, with per-step retries.

New

• Modular STT→LLM→TTS gateway — each stage streams end to end and swaps independently, with per-stage latency traced across the pipeline.

• Provider adapters — Deepgram and AssemblyAI for speech-to-text; ElevenLabs, Cartesia, PlayHT, Rime, LMNT, and Azure for speech — eleven adapters, all real and hot-swappable.

• Native Gemini Live — a single-model realtime path as an alternative to the composed pipeline.

• Barge-in — callers can interrupt the agent mid-sentence; it stops speaking and starts listening instantly.

New

• No-code agent builder — configure persona, voice, knowledge, tools, and guardrails in one place, with no prompt engineering required.

• Browser voice tester — talk to any draft agent live in the browser before it touches a phone line.

• Versioned publish & rollback — every publish is a version you can compare, promote, and roll back instantly.

New

• Multi-tenant workspaces with role-based access — owner, admin, operator, and viewer seats, backed by Firebase authentication and scoped server-side in the data layer.

• Agent projects & the studio shell — create, name, and organize voice and chat agents per workspace.

• Conversation store — every call and chat captured with transcript, metadata, and status.

• Activity feed and the first operator dashboard.